Healthcare AI Weekly Deep Dive

Anthropic is doing something strategically smart and strategically dangerous at the same time. The Optum partnership and the UST CarePath integration together represent a two-pronged distribution play: get Claude into payer back-office operations through Optum's advisory muscle, and get Claude into day-to-day insurer and provider workflows through UST's CarePath platform. CarePath is an operational platform that handles things like prior auth workflows, care management, and claims adjudication support. That is not ambient documentation. That is Claude making or informing real-time clinical and financial decisions. Here is the thing. Optum is the most vertically integrated organization in American healthcare. They own UnitedHealthcare. They own Optum Rx. They own Optum Health, which employs tens of thousands of physicians. And now they are integrating Claude. When a health system negotiates a value-based contract with UnitedHealthcare, the intelligence powering the other side of that negotiation may soon run on Claude. That is not paranoia. That is just following the org chart. For consultants, the immediate play is conflict mapping. Which of your clients have Optum contracts? Which have data-sharing agreements? Which are Optum Health medical groups? The Claude integration does not change the contractual relationships, but it does change who has analytical leverage. Systems that have not audited their Optum data flows in the last 18 months are flying blind. The second-order effect: Claude-powered claims adjudication makes denial logic faster and more consistent. Which means appealing those denials also needs to be faster and more consistent. Health systems that do not have their own AI-powered appeals tooling will fall further behind. The automation asymmetry is widening every quarter. UCSF's Converge accelerator and Pearl Health's $110M raise this same week signal that the counter-move is already underway: health systems and VBC players are building their own AI infrastructure specifically to avoid being on the wrong side of that asymmetry.

Two CMS actions dropped this week and the market is not fully processing how connected they are. First: CMS proposed phasing out traditional MIPS and expanding ACO participation, framing it as a shift toward prevention and accountable care. Second: CMS separately proposed major changes to how Medicare pays for clinical AI and remote patient monitoring. These are not separate policy tracks. They are the same bet placed twice. Here is the logic. If you push physicians into ACOs, you create immediate demand for population health AI: risk stratification, care gap closure, AWV automation, predictive admissions modeling. If you simultaneously update the payment framework for clinical AI and RPM, you create the reimbursement runway for those tools to generate real revenue. CMS is not just changing incentives. It is building the economic architecture for an AI-powered VBC ecosystem. The MIPS phase-out is the sharper edge. MIPS has been a checkbox exercise for most physician practices: report some quality measures, avoid the penalty, move on. ACO participation is a fundamentally different game. You are taking on financial risk for a defined population. You need to know who is going to the ER before they go. You need to close care gaps before the performance period ends. You need real-time data infrastructure. Most MIPS reporters do not have any of that. For consultants, this is a three-wave opportunity. Wave one: ACO feasibility and readiness assessments for physician groups currently in MIPS. Wave two: AI vendor selection for population health management in the new ACO framework. Wave three: performance optimization once the contracts are live. The firms that plant their flag in wave one will own the relationship through waves two and three. The RPM and clinical AI payment changes add a third dimension. If CMS creates cleaner reimbursement pathways for AI-assisted monitoring and clinical decision support, the ROI math on those investments changes overnight. Help your clients model the new numbers before the comment period closes.

The Heidi Health Pressure Points report is the most useful piece of data to land on a CMO or CNO's desk this week. Not because it is surprising, but because it makes the problem concrete with a number they can actually use. Eighty-three percent. That is not a fringe behavior. That is the dominant adoption pattern globally. Here is what that number actually means on the ground. A clinician at a 400-bed community hospital downloads an AI scribe app because her documentation time is killing her. She tells no one. She uses it for three months. Then she switches to a different one because her colleague recommended it. The hospital's IT department has no record of either tool. Their EHR vendor does not know. Their compliance team does not know. And if a note generated with the help of that tool is ever subpoenaed, the hospital cannot explain the workflow that produced it. This is what I call the shadow clinic. It is not one rogue physician. It is a majority of your clinical workforce using tools you did not select, did not vet, and did not train them on. The 88% figure on documentation burden being a critical issue explains why. When the administrative load is that heavy, people find relief wherever they can. You cannot govern your way out of a workflow problem. But you can build a governance framework that channels the behavior rather than trying to eliminate it. The consulting play is straightforward but it has to be done right. A blanket prohibition does not work. A lengthy policy document does not work. What works is an AI tool inventory process that takes two weeks, a tiered approval framework that moves fast for low-risk tools, and a clinician advisory board that gives frontline staff a sanctioned channel to surface what they are actually using. Do those three things and you go from liability exposure to controlled innovation in 90 days. The rural hospital story in this week's feed, about the NP leading AI implementation at a small system by killing the pilot carousel and building something real, is the right model. Clinical governance led by clinicians, not by IT. That is the pattern that sticks.

Two things happened in Washington this week on prior auth AI, and they point in opposite directions. The Senate blocked an effort to end CMS's Medicare AI prior authorization pilot, meaning the experiment has political cover to continue. And CMS ordered a corrective action plan against one of the AI vendors inside that same pilot, meaning it is already finding problems. Those two facts together tell you something important: the government wants AI in prior auth, but not unregulated AI. The corrective action is not a death knell for the category. It is a calibration signal. CMS is saying: we will let you operate, but not however you want. We do not yet know the specific vendor or the specific failure mode that triggered the corrective action. That matters enormously. If the AI was denying claims that should have been approved at a rate higher than human reviewers, that is a bias and accuracy problem. If it was approving claims that should have been denied, that is a fraud vulnerability. If it was operating without adequate explainability in its decision outputs, that is a transparency problem. Each of those failure modes has a different remediation pathway and a different liability profile. For health systems and payers that have already deployed AI prior auth tools, the question is not whether CMS is watching. They clearly are. The question is whether your vendor's tool has the same characteristics that triggered this corrective action. That requires your vendor to be transparent with you about how their model makes decisions, what its error rates look like by clinical category, and how it handles edge cases. If they cannot or will not answer those questions, that is your answer. The AMA's interoperability initiative on CPT code terminology mapping, announced this same week, is the structural foundation that makes AI prior auth less error-prone over time. If clinical terminology maps cleanly to CPT codes, the AI has better data to work with. That is a five-year fix. Your clients need to manage the near-term risk with what exists today.

Karan Singhal runs health AI at OpenAI and he is saying the right things to the right audience. Physicians are helping evaluate model performance. The vision is AI as a silent background companion for patients and clinicians. Models are getting better and you should build assuming they will keep improving. All of that is probably true. It is also exactly what OpenAI needs health systems to believe to close enterprise contracts. Here is the thing. 'Bet on the models getting better' is not a clinical governance framework. It is a vendor positioning statement. And the history of healthcare IT is littered with health systems that made technology bets based on vendor roadmaps that never materialized, or that materialized in ways that broke the workflows built on top of them. The specific risk with foundation models is different from the EHR vendor risk your clients already understand. When an EHR vendor releases an update, the clinical workflow usually breaks in predictable and visible ways. When a foundation model is updated, the output can shift in subtle ways that are much harder to detect. A diagnostic support tool that was well-calibrated for a specific set of clinical presentations may behave differently after a model update without anyone in the health system knowing the model changed. Singhal's framing of AI as a 'silent background companion' is actually the most honest part of the vision. The companion role is real and probably the right near-term deployment model. Silent background companion for documentation. Silent background companion for care gap identification. Silent background companion for risk stratification. None of those use cases require betting on future model capability. They require good integration with existing clinical data and careful validation against current model behavior. The challenge OpenAI is navigating is that healthcare procurement cycles are slow and contracts are long. They need health systems to sign now. 'Bet on models getting better' is the argument that makes signing now feel safe. It might even be true. But your clients should not sign on that bet without model-agnostic evaluation criteria written into their contracts.

Pearl Health has been quiet for a few months. A $110M Series C is not quiet. This is one of the more conviction-heavy funding rounds in VBC AI this year, and it landed the same week CMS proposed phasing out MIPS and expanding ACO participation. That timing is not accidental. Pearl's core play is building AI-powered infrastructure for primary care physicians operating in Medicare Shared Savings Program ACOs. Risk stratification, care gap identification, performance analytics, population health management. The tools that let a primary care practice actually function under a risk-bearing contract rather than just report quality measures and hope for the best. The MIPS phase-out is Pearl's best friend. Right now, most primary care practices can satisfy their CMS quality requirements by clicking through some MIPS attestation boxes once a year. When that option goes away and ACO participation becomes the primary alternative, the demand for Pearl's tools goes up sharply. The $110M is preparation for that demand surge. The risk is physician practice consolidation. Every time a large health system acquires an independent practice group, Pearl loses potential customers to the health system's existing population health infrastructure. The consolidation wave has been steady for a decade and is not slowing down. Pearl needs to convert practices before they get absorbed. For consultants advising health systems, the Pearl raise is a signal that the independent practice VBC ecosystem is being capitalized and will compete. Health systems that want to anchor their ACO networks with employed physicians will need to match what Pearl is offering to independent practices, or they will lose those referral relationships. The capability gap between what Pearl can offer a small independent practice and what a typical health system's population health tools offer that same practice is narrowing fast.

The KLAS Global HIT Trends 2026 report is useful not for the headline, which is expected, but for the tension buried in the data. AI is the top investment priority globally. And 37% of organizations are in introductory adoption phase. Those two facts together describe a market that is motivated to spend but not yet equipped to succeed. The US being the exception to the 'AI as top priority' pattern is interesting. In the US, the priority list is probably distorted by the weight of EHR optimization, revenue cycle pressure, and workforce burnout tooling. US health systems have a longer tail of immediate operational fires competing with AI for budget attention. International systems, particularly in markets with national health service infrastructure, may have cleaner data environments and simpler vendor landscapes that make AI adoption feel more tractable. The 37% introductory phase number is the one to watch. Organizations in introductory phase are organizations that are about to make vendor selection decisions at scale. They have declared AI a priority. Now they need to buy something. That buying cycle is happening right now and it is largely going to be driven by sales pitches from vendors rather than clinical evidence of efficacy. The 'pilot theater' problem is directly downstream from this dynamic. A health system in introductory phase wants to show leadership that it is doing AI. It launches three pilots. None of them are designed to measure outcomes in a way that would survive clinical scrutiny. The vendors get case studies. The health system gets dashboard screenshots. The patients see no difference. Eighteen months later, nobody can explain what changed. The antidote is a pre-purchase evaluation framework with outcome metrics baked in before the contract is signed. Not 'what does your product do' but 'how will we know in 12 months whether this worked and what does that data look like.' Most vendors cannot answer that question cleanly. That inability is actually your diagnostic.

The GE HealthCare and Catholic Health deal is big enough to generate press releases but the real story is in the structure, not the headline number. Five hundred million dollars over 10 years across cardiology, oncology, neurology, and women's health with 1,300-plus devices means GE is essentially becoming Catholic Health's exclusive technology spine for more than a decade. GE HealthCare is not purely a device company anymore. They have been investing heavily in AI-powered imaging analysis, clinical workflow software, and operational analytics. When you wire 1,300 devices into a single-vendor ecosystem, you are also wiring in that vendor's AI layer for every image interpreted, every workflow supported, and every piece of operational data generated on those devices. The upside for Catholic Health is real. Operational support at scale, predictable technology refresh cycles, integration that actually works because it is all from the same vendor. For a 40-site system, those are meaningful benefits. The downside is what you cannot see yet. When a better AI-powered diagnostic tool from a startup emerges and it does not integrate cleanly with GE's infrastructure, Catholic Health has to choose between the contractual path of least resistance and the clinically better option. That choice will come up repeatedly over a 10-year horizon. The contract terms on third-party interoperability will determine how painful those moments are. For consultants, this deal is a cautionary tale to put in front of clients who are being courted by large technology vendors with comprehensive platform pitches. The 10-year commitment sounds like stability. It is also a ceiling on your client's ability to respond to a rapidly changing AI landscape. Help them price the option value of flexibility before they sign it away.

Bunkerhill Health's Carebricks platform is answering a real question: who owns the AI agent layer in a health system? Right now, that layer belongs to whichever vendor the health system bought last. Epic owns it for Epic customers who use AI within the platform. Microsoft owns it for Azure-committed health systems. The ambient scribe vendor owns it for clinical documentation. Nobody owns it all, and nothing integrates. Bunkerhill's pitch is that health systems should build their own agents on shared infrastructure, the same way technology companies build microservices. You get a clinical scheduling agent. A prior auth agent. A care gap identification agent. They all run on Carebricks and they all talk to each other through standardized interfaces. The health system controls the logic. Bunkerhill provides the plumbing. The Khosla Ventures backing matters. Khosla does not do healthcare philanthropy. They did not put money into Bunkerhill because they like hospital systems. They put money in because they believe agentic AI infrastructure in healthcare is a platform-scale business. That is a significant vote of confidence in the category. The execution risk is the talent problem. Building AI agents on Carebricks requires engineers who understand both healthcare data and AI agent architecture. Most health systems have exactly zero of those people. The ones that do have technical talent have it concentrated in EHR optimization, not AI development. Bunkerhill's go-to-market has to solve that talent gap or the platform sits unused. For consulting firms, Carebricks is interesting for a different reason. If health systems start building their own agents, the implementation and integration work that comes with that is consulting territory. A health system that buys Carebricks and then needs help designing its clinical scheduling agent, integrating it with Epic, validating its outputs, and training its staff is a multi-year engagement. The platform creates the opportunity.

Lyric has been building what it calls Healthcare Decision Intelligence, a category that sounds like marketing until you unpack what Concert actually does. Concert translates machine-readable clinical policies into real-time claims workflows. That means taking the clinical criteria that determine whether a claim is covered and making them live inputs to the claims adjudication process rather than reference documents that human reviewers look up. Combined with Lyric's existing AI platform, that creates a closed loop where the clinical policy, the claims data, and the adjudication decision all live in the same system and update each other in real time. That is genuinely more sophisticated than how most payer claims systems work today. The downstream effect on providers is the part of this deal nobody is writing about. AI makes consistent denial logic cheap, which means fake-consistent denial logic also gets cheap. A payer that has wired clinical policy directly into its claims AI can generate denial rationales that are technically accurate to the policy language at a volume and speed that overwhelms most provider appeals processes. The provider appeals team is still doing manual review. The payer's denial engine is running in milliseconds. That asymmetry is already present in the market. This acquisition makes it worse. For provider clients, the response is the same it has been for the past two years: you need AI-powered appeals tooling, not just a bigger appeals team. The manual review model cannot keep pace with automated denial at scale. For payer clients, the Lyric acquisition signals that the category is consolidating around a few platforms that combine clinical policy intelligence with real-time claims AI. Being on the wrong platform three years from now is an expensive migration problem. Help them evaluate their current claims AI stack against where Lyric is going.

BCG publishing a healthcare AI framework is not news by itself. BCG publishes frameworks constantly. What matters is the category they chose: patient access centers. This is one of the highest-volume, highest-cost, highest-frustration administrative workflows in healthcare. It is also one of the categories where AI has the most credible near-term impact: scheduling automation, eligibility verification, prior auth initiation, insurance verification, call deflection. BCG choosing this category tells you where they see client buying intent. Health system COOs and CFOs are actively looking for ways to cut patient access center costs without cutting staff in a labor market where that is politically impossible. AI-powered patient access transformation is a story that sells in that environment: you get the cost improvement, you redeploy staff to higher-value work, you improve patient experience. It is one of the few AI narratives in healthcare where all three outcomes are credible and measurable. The competitive implication for other consulting firms is straightforward. BCG just put a flag in this territory with a published framework that will show up in client searches and conference presentations. The counter-move is not to publish your own patient access framework, because then you are playing BCG's game on BCG's turf. The counter-move is to own the adjacent territory with more clinical specificity. BCG's framework will be operationally competent and strategically sound but probably light on clinical workflow integration detail. That is where a healthcare-specialist firm can differentiate. For your clients, the BCG framework is useful as a benchmark. Read it, understand its assumptions, and help your clients identify where their specific patient access workflows diverge from the framework's generic model. That gap analysis is the conversation that generates real work. Chartis also surfaced this week through SAS's announcement of acing Chartis's 2026 fraud platform evaluation. Chartis publishing vendor evaluations in fraud and abuse AI is another form of market positioning. The firms that publish evaluations control the evaluation criteria, which means they also control what vendors optimize for and what health systems ask for in RFPs.